rsx: FIFO wake-up pause control

rpcs3/Emu/RSX/RSXFIFO.cpp

@@ -477,20 +477,21 @@ namespace rsx
 			fmt::throw_exception("Unexpected command 0x%x" HERE, cmd);
 		}
 
-		if (performance_counters.state != FIFO_state::running)
+		if (const auto state = performance_counters.state;
+			state != FIFO_state::running)
 		{
-			//Update performance counters with time spent in idle mode
-			performance_counters.idle_time += (get_system_time() - performance_counters.FIFO_idle_timestamp);
+			performance_counters.state = FIFO_state::running;
 
-			if (performance_counters.state == FIFO_state::spinning)
+			// Hack: Delay FIFO wake-up according to setting
+			// NOTE: The typical spin setup is a NOP followed by a jump-to-self
+			// NOTE: There is a small delay when the jump address is dynamically edited by cell
+			if (state != FIFO_state::nop)
 			{
-				//TODO: Properly simulate FIFO wake delay.
-				//NOTE: The typical spin setup is a NOP followed by a jump-to-self
-				//NOTE: There is a small delay when the jump address is dynamically edited by cell
-				busy_wait(3000);
+				fifo_wake_delay();
 			}
 
-			performance_counters.state = FIFO_state::running;
+			// Update performance counters with time spent in idle mode
+			performance_counters.idle_time += (get_system_time() - performance_counters.FIFO_idle_timestamp);
 		}
 
 		do

rpcs3/Emu/RSX/RSXThread.cpp

@@ -2273,6 +2273,61 @@ namespace rsx
 		}
 	}
 
+	void thread::fifo_wake_delay(u64 div)
+	{
+		// TODO: Nanoseconds accuracy
+		u64 remaining = g_cfg.video.driver_wakeup_delay;
+
+		if (!remaining)
+		{
+			return;
+		}
+
+		// Some cases do not need full delay
+		remaining = ::aligned_div(remaining, div);
+		const u64 until = get_system_time() + remaining;
+
+		while (true)
+		{
+#ifdef __linux__
+			// NOTE: Assumption that timer initialization has succeeded
+			u64 host_min_quantum = remaining <= 1000 ? 10 : 50;
+#else
+			// Host scheduler quantum for windows (worst case)
+			// NOTE: On ps3 this function has very high accuracy
+			constexpr u64 host_min_quantum = 500;
+#endif
+			if (remaining >= host_min_quantum)
+			{
+#ifdef __linux__
+				// Do not wait for the last quantum to avoid loss of accuracy
+				thread_ctrl::wait_for(remaining - ((remaining % host_min_quantum) + host_min_quantum), false);
+#else
+				// Wait on multiple of min quantum for large durations to avoid overloading low thread cpus
+				thread_ctrl::wait_for(remaining - (remaining % host_min_quantum), false);
+#endif
+			}
+			// TODO: Determine best value for yield delay
+			else if (remaining >= host_min_quantum / 2)
+			{
+				std::this_thread::yield();
+			}
+			else
+			{
+				busy_wait(100);
+			}
+
+			const u64 current = get_system_time();
+
+			if (current >= until)
+			{
+				break;
+			}
+
+			remaining = until - current;
+		}
+	}
+
 	u32 thread::get_fifo_cmd()
 	{
 		// Last fifo cmd for logging and utility

rpcs3/Emu/RSX/RSXThread.h

@@ -607,6 +607,7 @@ namespace rsx
 		atomic_t<bool> external_interrupt_ack{ false };
 		void flush_fifo();
 		void recover_fifo();
+		void fifo_wake_delay(u64 div = 1);
 		u32 get_fifo_cmd();
 
 		// Performance approximation counters

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -76,9 +76,23 @@ namespace rsx
 			const auto& sema = vm::_ref<atomic_be_t<u32>>(addr);
 
 			// TODO: Remove vblank semaphore hack
-			if (sema == arg || addr == rsx->device_addr + 0x30) return;
+			if (addr == rsx->device_addr + 0x30) return;
 
-			rsx->flush_fifo();
+			if (sema == arg)
+			{
+				// Flip semaphore doesnt need wake-up delay
+				if (addr != rsx->label_addr + 0x10)
+				{
+					rsx->flush_fifo();
+					rsx->fifo_wake_delay(2);
+				}
+
+				return;
+			}
+			else
+			{
+				rsx->flush_fifo();
+			}
 
 			u64 start = get_system_time();
 			while (sema != arg)
@@ -113,6 +127,7 @@ namespace rsx
 				std::this_thread::yield();
 			}
 
+			rsx->fifo_wake_delay();
 			rsx->performance_counters.idle_time += (get_system_time() - start);
 		}
 

rpcs3/Emu/System.h

@@ -502,7 +502,8 @@ struct cfg_root : cfg::node
 		cfg::_int<50, 800> resolution_scale_percent{this, "Resolution Scale", 100};
 		cfg::_int<0, 16> anisotropic_level_override{this, "Anisotropic Filter Override", 0};
 		cfg::_int<1, 1024> min_scalable_dimension{this, "Minimum Scalable Dimension", 16};
-		cfg::_int<0, 30000000> driver_recovery_timeout{this, "Driver Recovery Timeout", 1000000};
+		cfg::_int<0, 30000000> driver_recovery_timeout{this, "Driver Recovery Timeout", 1000000, true};
+		cfg::_int<0, 16667> driver_wakeup_delay{this, "Driver Wake-Up Delay", 1, true};
 		cfg::_int<1, 500> vblank_rate{this, "Vblank Rate", 60}; // Changing this from 60 may affect game speed in unexpected ways
 
 		struct node_vk : cfg::node